DEVELOPMENT OF AI ALGORITHM FOR SPEECH QUALITY AND INTELLIGIBILITY ASSESSMENT UNDER NOISY CONDITION

Abstract

Clean speech signal as reference required for most objective speech intelligibility assessment metrics called intrusive methods for speech assessment. This limits the application of these intrusive metrics in real world situation. So deep learning based models are used in this thesis which does not require clean speech as reference signal, these methods are called non-intrusive methods. The input of deep learning model is fourier transform of degraded signal and output is made to predict PESQ and STOI metrics for quality and intelligibility respectively. The experimental results show estimated PESQ and STOI have good correlation with actual PESQ & STOI. The MSE values for PESQ and STOI are 0.0143 and 0.003. The MAE values are 0.118 and 0.034 respectively. These results show that the deep learning model is able to predict objective scores without any reference signal.